Golden Finance reports that today we are simultaneously launching two official model versions: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. DeepSeek-V3.2 is our first model to integrate thinking into tool usage, and it simultaneously supports both thinking mode and non-thinking mode for tool calls. We proposed a large-scale agent training data synthesis method, constructing a vast number of "hard to answer, easy to verify" reinforcement learning tasks (1800+ environments, 85,000+ complex instructions), significantly enhancing the model's generalization ability. (DeepSeek)

PANews news on November 27th reports that DeepSeek has launched a new mathematical reasoning model, DeepSeekMath-V2, which uses a self-verifying training framework. This model is built on DeepSeek-V3.2-Exp-Base and automatically reviews generated mathematical proofs through an LLM verifier, continuously optimizing performance using high-difficulty samples.

BlockBeats news, on November 5th, according to on-chain AI analysis tool CoinBob (@CoinbobAIbot) monitoring, the actual positions of the six mainstream AI large models in the "AI Cryptocurrency Trading Competition" season one were uniformly liquidated by the artificial intelligence research laboratory nof1 at 2 AM today, resulting in losses for all. nof1 started a large model trading test called Alpha Arena on October 18, involving six mainstream AI large models (GPT-5, Gemini 2.5 Pro, Grok-4, Claude Sonnet 4.5, DeepSeek V3.1, Qwen3 Max).

The first nof1 AI model trading competition has concluded, with Qwen3 Max ranking first with a return of 22.3% and a total profit and loss of $2,232, followed by DeepSeek Chat V3.1 in second place. Other models experienced significant losses, with the highest loss reaching 62.66%.

Odaily News nof1 founder Jay A announced on X that the Alpha Arena Season 1 competition officially ended this morning, with Qwen 3 MAX surpassing Deepseek Chat V3.1 at the last moment to become the AI model with the highest final profit.

The market has been sideways for two days. The AI model's holdings show that both DeepSeek and Qwen3 have unrealized gains, while Grok and Gemini, which have long held short orders, are suffering significant losses. The rankings of the holdings for each model are DeepSeek, Qwen3, and Claude, etc.

With the fluctuation of the crypto market, the performance of the six major AI trading models has declined, with only DeepSeek and QWEN3 showing unrealized gains, while the other models are all experiencing unrealized losses. Among them, DeepSeek's unrealized gain is the highest at 5,339.29 USDC.

As the market falls, several large model accounts participating in AlphaZero AI trading have shown significant drawdowns, with the DeepSeek account falling by 32.3%, Qwen3 down by 29.8%, and GPT5 losing 72.6%, mainly due to high leverage and loose stop loss strategies.

Recently, the market has warmed up, and the returns of the six major AI models have obviously diverged, with DeepSeek leading at a return of 125.92%, followed by Qwen3. DeepSeek achieved success by adopting a 10x leverage long position strategy on six mainstream tokens, with significant profits from long orders on BTC and SOL. The overall account balance ranking is DeepSeek, Qwen3, Claude, Grok, Gemini, and GPT5.

As the price of Bitcoin warms up to $114,000, the trading account funds of the six major AI models have recovered. QWEN3 performed the best, with a return rate close to 100%, with a total account amount of $19,571.36. DeepSeek follows closely behind, with a total amount of $18,292.30. Other models like Claude and Grok have also recovered to around $10,000 levels.

BlockBeats news, on October 23, according to official information, Qwen3 Max surpassed DeepSeek in the AI Cryptocurrency Trading competition held by the artificial intelligence research lab nof1, currently ranking first with a balance of $12,100 (+21%). Qwen3 Max8 explained that while its portfolio rose by 21%, its Sharpe ratio is negative. Therefore, it will maintain a 20x Bitcoin long positions with a take profit target set at $112,253.96 and a stop loss at $105,877.7. The current cash balance is only $96.8, so no new trades will be conducted for now. According to on-chain AI analysis tool CoinBob (@CoinbobAIbot), the Qwen3 Max8 Address was tracked and analyzed on October 21 at $107,993.

Odaily reported that in the "nof1 AI Automated Trading Competition", Deepseek Chat V3.1 ranked first with $11,085, followed closely by Grok 4 and Qwen3 Max, while Claude Sonnet 4.5, Gemini 2.5 Pro, and GPT-5 assets experienced varying degrees of decline.

According to Foresight News, data from the Polymarket official website shows that the probability of betting on "the winner of the NOF1 AI trading competition" for DeepSeek is 36%, for xAI is 29%, and for Anthropic is 28%.

Zhao Changpeng commented on the 'nof1 AI Automated Trading Competition', stating that DeepSeek has performed outstandingly in AI trading. He believes that unique trading strategies are more effective and that if people use the same AI collectively, it may lead to price fluctuations. He predicts that research on AI trading will increase in the future, and the volume will also rise.

On October 20, news came that the AI research lab nof1 began testing the large model trading Alpha Arena, with 6 mainstream AI large models participating in the test. So far, DeepSeek and Grok have achieved returns exceeding 14%, while Gemini 2.5 Pro has incurred a loss of 42.57%.

On September 29, Jin10 reported that Cambrian announced in its public account that on September 29, 2025, Cambrian had synchronized the adaptation of the latest model DeepSeek-V3.2-Exp from Deep Seeking Company and open-sourced the large model inference engine vLLM-MLU source code. Regarding the new model architecture DeepSeek-V3.2-Exp, Cambrian achieved rapid adaptation through the Triton operator development, utilized the BangC fusion operator development for extreme performance optimization, and based on the parallel strategy of computation and communication, achieved industry-leading computational efficiency levels once again.

Jin10 data reported on September 29th that DeepSeek announced that the official App, web version, and mini program have all been updated to DeepSeek-V3.2-Exp. DeepSeek introduced that due to a significant drop in the service cost of the new model, the official API price has also been adjusted accordingly, with the new prices taking effect immediately. Under the new pricing policy, the cost for developers to call the DeepSeek API will be reduced by more than 50%.

Jin10 reported on September 29 that AI practitioners discovered that the DeepSeek-V3.2 new model had been uploaded to the community platform HuggingFace, and was subsequently deleted. Additionally, it is understood that the new GLM-4.6 model from Zhichu will be released soon and is currently accessible via API interface.

DeepSeek announced tonight that its online model has been upgraded, with the current version number DeepSeek-V3.1-Terminus, which includes two versions: the thinking model and the non-thinking model, both with a context length of 128k, and has supported users for online experience. Among them, the default output length of the non-thinking model is 4K, with a maximum of 8K, and the default output length of the thinking model is 32K, maximum.

PANews, September 18th news, DeepSeek released an official statement saying that recently, criminals have impersonated the company's officials or employees, committing fraud in the name of "Computing Power Leasing" and "Equity Financing", forging work badges, business licenses and other materials to charge users fees. The company emphasizes that it has never required users to make payments to personal accounts or unofficial accounts. Any private transfer behavior constitutes fraud, and the company will pursue legal actions against those who illegally use the company's name. Users should obtain information through the official website and officially certified accounts. The official webpage and App products are currently free.

Golden Finance reports that market news: DeepSeek aims to launch its AI agent (AI AGENT) before the end of the year.

DeepSeek officially released version V3.1, which includes a hybrid inference architecture, improved thinking efficiency, and stronger Agent capabilities. Users can freely switch between thinking mode and non-thinking mode via the "Depth Thinking" button.

Golden Finance reported that early this morning, DeepSeek released the Deepseek V3.1-Base version. (AIGC Open Community)

Jin10 Data reported on August 13 that recently, the market has once again heard rumors about the release of the next-generation large model DeepSeek-R2, with an expected time window from August 15 to 30. Sources close to DeepSeek stated that the news is untrue and confirmed that there are no plans to release DeepSeek-R2 within August.

PANews August 11 news, the official status page shows that DeepSeek's web and API services experienced a major failure today. The official team has identified the issue and is gradually restoring the services.

Gate News bot message, the privacy-preserving computing network Nillion has completed its Mainnet Phase 0 upgrade for the first time. According to the latest upgrade, developers gain the ability to run AI models such as Llama and DeepSeek in TEE. This upgrade allows developers to store and query encrypted data on a decentralized Node cluster without needing to rely on node operators.

On June 26, Jin10 reported that Cambrian expressed on the interactive platform that the products developed by the company are general-purpose intelligent chips that can adapt to mainstream Open Source models, including DeepSeek.

Gate News bot message, MiniMax announced on June 17 that it will release important updates for five consecutive days. Today's first release is the open source inference model MiniMax-M1. According to the official report, MiniMax-M1 has benchmark tests that rival open source models like DeepSeek-R1 and Qwen3, approaching the most advanced models overseas. The official blog also mentioned that based on two major technological innovations, the training process of MiniMax-M1 was efficiently "beyond expectations," completing the reinforcement learning training phase in just 3 weeks with 512 H800 GPUs, and the computing power rental cost was only $534,700.

PANews, May 29 - According to official news from DeepSeek, the new model version DeepSeek-R1-0528 has been launched, bringing improvements in benchmark performance, enhanced front-end capabilities, and a drop in hallucination rates, along with new support for JSON output and function calling features. The API interface remains unchanged, and the model weights have been released as Open Source. Users can immediately visit chat.deepseek.com to experience it.

PANews, May 21st news, Nvidia CEO Jen-Hsun Huang: DeepSeek has increased the demand for artificial intelligence computing by 100 to 1000 times.

Odaily News Microsoft has announced that it is adding Musk's xAI model to its artificial intelligence marketplace, which includes Grok 3, set to launch earlier in 2025. Microsoft stated that users of its Azure cloud services can access over 1,900 AI models, including those from close partners Microsoft, OpenAI, Meta Platforms, and DeepSeek. (Jin10)

Jin10 reported on May 19 that Nvidia CEO Jen-Hsun Huang stated that the DeepSeek-R1 demonstrated groundbreaking performance improvements, with its computing power increased fourfold compared to leading competitors like the H100. The DeepSeek-R1 represents an important innovation in the global artificial intelligence industry, influencing researchers' approaches to the study of AI and reasoning, and opening up new research areas.

Jin10 reported on May 13 that the DeepSeek service status page has been updated, stating that the APP and WEB chat functions have returned to normal, although there is a chance of failure in retrieving chat history.

Jin10 data reported on April 21st that iFLYTEK's Depth inference large model—Xunfei Spark X1 has received a new upgrade. The upgraded Spark X1 shows significant improvements in general tasks such as mathematics, coding, logical reasoning, text generation, language understanding, and knowledge Q&A. Despite having model parameters an order of magnitude smaller than comparable models in the industry, its overall performance is on par with OpenAI o1 and DeepSeek R1.

On April 14, Jin10 reported that Citigroup stock strategists downgraded their rating on the U.S. stock market, stating that recent events such as DeepSeek, Europe's fiscal stance, and trade tensions have reinforced their view on diversifying investments outside the U.S. market. "From a GDP and EPS perspective, the drivers of the 'exceptionalism' are fading," said Beata Manthey and other strategists at the firm, downgrading the U.S. stock market rating from overweight to neutral. Citigroup stated that the U.S. market remains relatively expensive, and the trend of downward EPS expectations is intensifying.

Gate.io News bot message, Zhu Su recently published a post sharing his market observations in Hong Kong. According to his observations, the Art Basel art fair did not involve NFT content, and the market discussion on memes has dropped. Investment banks are promoting the theme of "China's Technology Revival", focusing on projects like DeepSeek, and paying attention to valuation gaps and innovation trends. Data shows that southbound capital continues to flow in, and the price of mainboard shell companies has reached approximately 200 million HKD. Family offices are closely monitoring Saylor's movements and considering adjustments to their company structure. The market generally positions Bitcoin as "digital gold," but maintains a conservative attitude towards the prospects of crypto technology and venture capital. Some investors express concerns about regulated custody and exchange investment projects. Source: Wu Says

Jin10 reported on March 22 that Nezha 2 has extended its release to April 30, and DeepSeek optimistically predicts that the global box office for the film may reach 16 billion.

On March 21, Jinshi Data reported that Futu announced at the "Futu Investment Expo" held at the Hong Kong Convention and Exhibition Centre on March 21 that it will officially integrate DeepSeek-R1 and introduce multiple upgrades of artificial intelligence ("AI") product features to empower retail investors with top investment vehicles, further realizing the vision of inclusive finance. Futu also announced a series of the latest product features for 2025, including Niu Niu AI, a new "Opportunities" smart investment theme page, Options strategies, algorithmic trading, and more.

On March 20, Jinshi Data reported that StarNet RuiJie launched the DeepSeek enterprise private Computing Power integrated machine, providing a plug-and-play environment for large model fine-tuning and inference, empowering intelligent transformation across various industries with strong AI capabilities.

On March 20, Jinshi Data reported that the API service and web chat service of DeepSeek R1 had previously been unavailable, but it has now returned to normal.

On March 20, Jinshi Data reported that on March 19 local time, NVIDIA (NVDA.O) CEO Jen-Hsun Huang stated at the GTC conference that new artificial intelligence models capable of producing more complex answers will only increase the demand for computing infrastructure. Jen-Hsun Huang stated that concerns from the outside world about the DeekSeek AI model R1 needing fewer chips and more powerful servers are incorrect. "The understanding of R1 from the outside is completely wrong; in fact, this model has a much higher demand for computing power."

PANews March 19 news, Nvidia CEO Jen-Hsun Huang: The understanding of DeepSeek R1 is completely wrong. The demand for computation remains very high, and inference artificial intelligence has increased the demand for computation.

On March 19, Jinshi Data reported that CMB Securities stated on the interactive platform that the company has deployed the DeepSeek series of models, which are currently mainly integrated with the company's office communication software to support employees' daily office needs. In the future, the company will comprehensively use R1 class strong reasoning models for business intelligent upgrade, focusing on landing exploration around core businesses such as intelligent investment consulting, intelligent investment, and intelligent investment research.

Golden Ten Data, March 17, today, 010000 announced the launch of a one-stop platform for Magic Enterprise Large Model to deploy customized solutions for enterprise-level DeepSeek.

SIPG recently launched the "MaaSAI Assistant", which uses artificial intelligence technology to analyze port operation data in real time and improve port service capabilities. In the future, we will continue to promote the in-depth integration of AI technology and business, create specialized AI applications, and build competitive advantages.

Jinshi data news on March 13th, Zhongguancun Kejin released DeZhu's large model platform 2.5 and DeZhu's intelligent customer service 4.0. Among them, DeZhu's large model platform 2.5 has now accessed DeepSeek, covering Computing Power, data, models, intelligent entities, and other full-chain large model development and application capabilities. DeZhu's intelligent customer service 4.0 also has the full-stack capability of intelligent customer service products, covering a full set of intelligent customer service products including full-media contact centers, voice robots, and text robots. Zhongguancun Kejin stated that it is expected that by 2025, over 90% of enterprises will integrate large model Depth into their core business.

Recently, WeBank held its first "Media Open Day", showcasing its achievements over the past decade. It has served over 420 million individual customers and over 5.8 million corporate customers, with the annual IT operation and maintenance cost per account as low as 2 yuan, and the peak daily transaction volume reaching 1.4 billion. WeBank is committed to building an AI-native paradigm and has already established capabilities in AI infrastructure, applications, and governance. Its self-developed AI engineering platform supports the private deployment of mainstream Open Source large models such as DeepSeek.